Accurate and early yield predictions through advanced statistical modelling
To determine if more accurate crop forecasting tools can be developed, using machine learning methods. The output will be a web-based resource for growers, with a target yield prediction accuracy of ± 5%. The tool will comprise both a crop forecasting model and a berry growth model, which will be able to be used either separately or in tandem.
Yield estimation tools are already available but their accuracy varies widely, from ± 20% to at best, ± 5% immediately before harvest. They also require extensive sampling in vineyards after berries have gone through veraison. Previous work has identified the factors involved in yield but the impact of their interactions is not well understood. Modelling this level of complexity is not trivial and has not been undertaken in the past, nor has the development of a yield estimation tool from an understanding of these interactions been previously attempted.
This is a desktop project using machine learning methods (e.g. Artificial Neural Networks) to interpret the interaction of the multiple variables that can impact on yield. Over 20 years of vineyard data from Treasury Wine Estates will be provided as will data from Casella, McWilliams and Brown Brothers. The aim is to include data in the model that growers will be likely to collect for other reasons, so that the labour required to use the yield prediction tool will be minimised. A browser interface will be developed to allow the tool to be made more accessible for use by growers.
A new yield prediction tool will be created that will be superior to current methods in three ways:
1) It will incorporate weather data interactions with vine characteristics at key developmental stages.
2) It will produce more accurate predictions from fewer measurements in the vineyard.
3) Crop forecasts will be available as early as winter pruning, and will be continually updated with increasing precision as the season progresses.